Reliability-Aware Deep Learning Framework for Chemical Genotoxicity Prediction with Uncertainty Quantification.

Area of Science:

• Computational toxicology
• Drug development
• Chemical safety assessment

Background:

• Genotoxicity testing is vital but faces challenges with traditional experimental methods.
• Existing computational models often overlook data quality and predictive uncertainty.
• There is a need for more reliable and transparent genotoxicity prediction tools.

Purpose of the Study:

• To develop a reliability-aware computational framework for genotoxicity prediction.
• To address data heterogeneity and predictive uncertainty in genotoxicity assessments.
• To improve the efficiency and ethical considerations in drug development and chemical safety.

Main Methods:

• Utilized a curated dataset of 8,389 compounds with experimental reliability tiers.
• Employed a two-step hierarchical learning strategy with message-passing neural networks and conventional machine learning models (Random Forest, SVM).
• Integrated conformal prediction for quantifying predictive uncertainty and providing coverage guarantees.

Main Results:

• Random Forest and RBF-kernel SVM achieved high predictive performance (AUC 0.8613 and 0.8582).
• Conformal prediction demonstrated 90.7% empirical coverage and identified ambiguous predictions.
• The framework successfully incorporated data reliability and uncertainty quantification.

Conclusions:

• The proposed framework offers a more transparent and uncertainty-aware approach to genotoxicity prediction.
• Accounting for data reliability and uncertainty is crucial for robust computational toxicology.
• This method can aid in more efficient and ethical drug development and chemical safety evaluations.